#include "ccv.h"

#include "ccv_internal.h"

#include "nnc/ccv_nnc.h"

#include "nnc/ccv_nnc_easy.h"

#include "nnc/ccv_nnc_internal.h"

#ifdef USE_OPENMP

#include <omp.h>

#endif

#ifdef USE_DISPATCH

#include <dispatch/dispatch.h>

#endif

// Shared methods.

#include "../_ccv_nnc_cpu_ref.h"

void _ccv_nnc_reduce_mean_forw_cpu_ref(ccv_nnc_tensor_view_t* const a, ccv_nnc_tensor_view_t* const b)

{

  assert(ccv_nnc_tensor_nd(a->info.dim) <= CCV_NNC_MAX_DIM + 2);

  assert(ccv_nnc_tensor_nd(b->info.dim) <= CCV_NNC_MAX_DIM + 2);

  // Assuming this is float 32.

  int adim[CCV_NNC_MAX_DIM_ALLOC];

  int bdim[CCV_NNC_MAX_DIM_ALLOC];

  ccv_nnc_tensor_view_get_dim(a, adim);

  ccv_nnc_tensor_view_get_dim(b, bdim);

  assert(ccv_nnc_tensor_view_check_broadcast_dim(b, adim));

  int astride[CCV_NNC_MAX_DIM_ALLOC];

  int bstride[CCV_NNC_MAX_DIM_ALLOC];

  assert(CCV_NNC_MAX_DIM == 2); // Need to change this logic for CCV_NNC_MAX_DIM == other number.

  ccv_nnc_tensor_view_get_stride(a, astride);

  ccv_nnc_tensor_view_get_stride(b, bstride);

  int i[CCV_NNC_MAX_DIM + 2];

  int x;

  ccv_nnc_tensor_zero(b);

  float* const ap = a->data.f32;

  float* const bp = b->data.f32;

  const float scale = (float)ccv_nnc_tensor_count(b->info) / (float)ccv_nnc_tensor_count(a->info);

  // Non-optimal case, need to do skip if needed.

  for (i[0] = 0; i[0] < adim[0]; 
i[0]++5
)

  {

    float* const ap0 = ap + i[0] * astride[0];

    float* const bp0 = bdim[0] == 1 ? bp : 
bp + i[0] * bstride[0]0
;

    for (i[1] = 0; i[1] < adim[1]; 
i[1]++5
)

    {

      float* ap1 = ap0 + i[1] * astride[1];

      float* const bp1 = bdim[1] == 1 ? bp0 : 
bp0 + i[1] * bstride[1]0
;

      for (i[2] = 0; i[2] < adim[2]; 
i[2]++14
)

      {

        float* const bp2 = bdim[2] == 1 ? 
bp14
 : 
bp1 + i[2] * bstride[2]10
;

        if (bdim[3] == 1)

          
for (x = 0; 10
x < adim[3]; 
x++18
)

            bp2[0] += ap1[x] * scale;

        else

          
for (x = 0; 4
x < adim[3]; 
x++12
)

            bp2[x] += ap1[x] * scale;

        ap1 += astride[2];

      }

    }

  }

}

static int _ccv_nnc_reduce_mean_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)

{

  assert(input_size == 1);

  ccv_nnc_tensor_view_t* const a = (ccv_nnc_tensor_view_t*)inputs[0];

  ccv_nnc_tensor_view_t* const b = (ccv_nnc_tensor_view_t*)outputs[0];

  _ccv_nnc_reduce_mean_forw_cpu_ref(a, b);

  return CCV_NNC_EXEC_SUCCESS;

}

static int _ccv_nnc_reduce_mean_back(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)

{

  if (inputs[0] == 0)

  {

    int i;

    ssize_t dims = 1;

    for (i = 0; i < cmd.info.reduce.count; i++)

      dims *= outputs[0]->info.dim[cmd.info.reduce.axis[i]];

    const float scale = 1.0 / (float)dims;

    _ccv_nnc_tensor_set_cpu_ref_f32((ccv_nnc_tensor_view_t*)outputs[0], scale);

    return CCV_NNC_EXEC_SUCCESS;

  }

  ccv_nnc_tensor_view_t* const a = (ccv_nnc_tensor_view_t*)outputs[0];

  ccv_nnc_tensor_view_t* const b = (ccv_nnc_tensor_view_t*)inputs[0];

  assert(ccv_nnc_tensor_nd(a->info.dim) <= CCV_NNC_MAX_DIM + 2);

  assert(ccv_nnc_tensor_nd(b->info.dim) <= CCV_NNC_MAX_DIM + 2);

  // Assuming this is float 32.

  int adim[CCV_NNC_MAX_DIM_ALLOC];

  int bdim[CCV_NNC_MAX_DIM_ALLOC];

  ccv_nnc_tensor_view_get_dim(a, adim);

  ccv_nnc_tensor_view_get_dim(b, bdim);

  int astride[CCV_NNC_MAX_DIM_ALLOC];

  int bstride[CCV_NNC_MAX_DIM_ALLOC];

  assert(CCV_NNC_MAX_DIM == 2); // Need to change this logic for CCV_NNC_MAX_DIM == other number.

  ccv_nnc_tensor_view_get_stride(a, astride);

  ccv_nnc_tensor_view_get_stride(b, bstride);

  int i[CCV_NNC_MAX_DIM + 2];

  int x;

  float* const ap = a->data.f32;

  float* const bp = b->data.f32;

  const float scale = (float)ccv_nnc_tensor_count(b->info) / (float)ccv_nnc_tensor_count(a->info);

  // Non-optimal case, need to do skip if needed.

  for (i[0] = 0; i[0] < adim[0]; 
i[0]++1
)

  {

    float* const ap0 = ap + i[0] * astride[0];

    float* const bp0 = bdim[0] == 1 ? bp : 
bp + i[0] * bstride[0]0
;

    for (i[1] = 0; i[1] < adim[1]; 
i[1]++1
)

    {

      float* ap1 = ap0 + i[1] * astride[1];

      float* const bp1 = bdim[1] == 1 ? bp0 : 
bp0 + i[1] * bstride[1]0
;

      for (i[2] = 0; i[2] < adim[2]; 
i[2]++2
)

      {

        float* const bp2 = bdim[2] == 1 ? bp1 : 
bp1 + i[2] * bstride[2]0
;

        if (bdim[3] == 1)

          for (x = 0; x < adim[3]; x++)

            ap1[x] = bp2[0] * scale;

        else

          
for (x = 0; 2
x < adim[3]; 
x++6
)

            ap1[x] = bp2[x] * scale;

        ap1 += astride[2];

      }

    }

  }

  return CCV_NNC_EXEC_SUCCESS;

}

REGISTER_COMMAND_BACKEND(CCV_NNC_REDUCE_MEAN_FORWARD, CCV_NNC_BACKEND_CPU_REF)(ccv_nnc_cmd_backend_registry_t* const registry)

{

  registry->tensor_formats = CCV_TENSOR_FORMAT_NHWC | CCV_TENSOR_FORMAT_NCHW | CCV_TENSOR_FORMAT_CHWN;

  registry->tensor_datatypes = CCV_32F;

  registry->tensor_memory = CCV_TENSOR_CPU_MEMORY;

  registry->algorithms = 1;

  registry->exec = _ccv_nnc_reduce_mean_forw;

}

REGISTER_COMMAND_BACKEND(CCV_NNC_REDUCE_MEAN_BACKWARD, CCV_NNC_BACKEND_CPU_REF)(ccv_nnc_cmd_backend_registry_t* const registry)

{

  registry->tensor_formats = CCV_TENSOR_FORMAT_NHWC | CCV_TENSOR_FORMAT_NCHW | CCV_TENSOR_FORMAT_CHWN;

  registry->tensor_datatypes = CCV_32F;

  registry->tensor_memory = CCV_TENSOR_CPU_MEMORY;

  registry->algorithms = 1;

  registry->exec = _ccv_nnc_reduce_mean_back;

}